Design And Implementation Of Early Gastric Cancer Detection System Based On Deep Learning

The 21st century is an era of intelligence and information technology.Artificial intelligence is playing an important role all the time in the intelligent transformation of all walks of life and people’s daily necessities.As technology advances,AI(Artificial Intelligence)is gradually being integrated into the medical industry.In addition to the digitalization and intellectualization of the medical platform,intelligent algorithms,big data analysis and other methods can be used to mine and utilize the information in a large amount of medical data to form a certain diagnosis,which can play a role in assisting doctors in treatment.Gastric cancer has a very high incidence and mortality in China.However,due to its early clinical manifestations similar to gastritis,it is hard to attract the attention of patients,and most of them have developed into advanced gastric cancer when diagnosed.However,most early gastric cancer can be treated by endoscopy,and the 5-year survival rate is more than 90%.Using computer vision to detect and further analyze the early gastric cancer lesions in gastroscopic images,and visualization of the results through web-based early gastric cancer detection system can assist doctors in diagnosis.The main research work of this paper is as follows:1.Acquisition,pretreatment and enhancement of gastroscopic image dataset.The gastroscopic images of patients were obtained from the cooperative hospitals,and the doctors annotated the gastroscopic images according to the biopsy or ESD pathological results as the gold standard.Then,a new data enhancement method combining the traditional method and improved DCGAN was proposed to enhance the data of gastroscopy image data,alleviating the imbalance between datasets and the problems of small data amount.The effectiveness and necessity of data enhancement were proved by comparative experiments.2.Mask R-CNN+BiFPN of improved FPN(Feature Pyramid Network)was proposed to detect early gastric cancer lesions in gastroscopic images.In the design and implementation of the detection network for early gastric cancer lesions,the detection effect of basic network Faster R-CNN and Mask R-CNN was firstly compared,and the Mask R-CNN network with better performance was selected.Then on this basis,aiming at the problem of insufficient feature fusion ability of FPN network,we proposed the Mask R-CNN+BiFPN network by referring to BiFPN network structure and combining with ResNet50 feature extraction network.The experimental results show that the improved network can significantly improve the detection effect of gastric cancer lesions.3.The texture feature extraction of gastric cancer image and the segmentation of early gastric cancer lesions based on improved U-net were performed.After the detection of early gastric cancer lesions in gastroscopic images,in order to make the detection results more explanatory and provide doctors with more diagnostic reference,two methods of texture feature extraction of gastroscopic images and segmentation of early gastric cancer lesions were used to further process the gastroscopic images.Gastric cancer lesion segmentation network is an improvement of U-net.By adjusting the loss function and adding the attention mechanism to strengthen the feature extraction network,the segmentation effect of the network is improved.4.Design and implementation of web-based early gastric cancer detection system.After completing the training and comparative experiments of all models,the web-based early gastric cancer detection system was developed and designed,with python-flask as the back-end framework and Vue as the front-end framework.The three functional modules of detection of early gastric cancer lesions,texture feature extraction of gastroscopic image and segmentation of early gastric cancer lesions are integrated into the system,which can be selected and used by doctors according to their own needs.After uploading gastroscopic images,the results of system detection will be displayed in the browser.In this paper,the method of deep learning is adopted to detect the early gastric cancer lesions in gastroscopic images,and at the same time,the texture feature extraction of gastroscopic images and the segmentation of gastric cancer lesions are further carried out.Finally,the results are displayed on the webpage in the form of images,so as to achieve the purpose of assisting doctors in diagnosis.Doctors can easily and quickly obtain the detection results of early gastric cancer in gastroscopic images,completing the combination of intelligent,digital and medical images.